# Copy to project root: cp .env.example .env

# LLM provider: rapid_mlx | ollama | openai_compatible
LLM_PROVIDER=rapid_mlx

# Rapid-MLX endpoint (see macos/.env.example for HF auth)
RAPID_MLX_HOST=127.0.0.1
RAPID_MLX_PORT=8080
RAPID_MLX_BASE_URL=http://127.0.0.1:8080/v1
RAPID_MLX_MODEL=gemma-4-e4b-4bit
RAPID_MLX_MODEL_ENDPOINTS=mlx-community/Qwen3.5-4B-MLX-4bit=http://127.0.0.1:8081/v1,LiquidAI/LFM2-1.2B-RAG=http://127.0.0.1:8082/v1
RAPID_MLX_API_KEY=not-needed

# Optional online OpenAI-compatible Gemma 4 12B endpoint
# GEMMA_ONLINE_BASE_URL=http://matrix:PORT/v1
# GEMMA_ONLINE_API_KEY=replace-me

# Ollama
OLLAMA_BASE_URL=http://localhost:11434
OLLAMA_MODEL=qwen3.5:4b

# OpenAI-compatible (LM Studio, vLLM, etc.)
OPENAI_BASE_URL=http://localhost:1234/v1
OPENAI_API_KEY=not-needed
OPENAI_MODEL=local-model

# Cursor models (requires `agent` CLI on PATH; appear alongside LLM_PROVIDER)
# Full CLI: write/shell tools, --force, sandbox disabled (not ask/plan read-only)
# CURSOR_API_KEY=
# CURSOR_MODELS=auto,cursor-grok-4.5-high-fast
# CURSOR_AGENT_BIN=agent

# ChatGPT account (OpenAI OAuth via openai-oauth local proxy; Settings → Add model)
# OPENAI_OAUTH_HOST=127.0.0.1
# OPENAI_OAUTH_PORT=10531
# OPENAI_OAUTH_BIN=npx --yes openai-oauth@2
# OPENAI_OAUTH_AUTH_FILE=
# RS2_OPENAI_OAUTH_MODELS_PATH=

# Optional path for Settings → Add model (default: ~/.rs2/custom-models.json)
# RS2_CUSTOM_MODELS_PATH=

# Chat server
HOST=127.0.0.1
PORT=8000
# Optional path overrides (auto-detected from workspace / cwd when unset)
# RS2_ROOT=
# RS2_WEB_DIR=

# Comma-separated browser origins (use * for local dev only)
# CORS_ORIGINS=*

# Math evaluation timeout in seconds (SymPy path)
# MATH_EVAL_TIMEOUT_SECONDS=5.0

# Context limits (also exposed via GET /api/health → context_limits)
# CONTEXT_MAX_MESSAGES=100
# CONTEXT_MAX_CHARS=32000
# CONTEXT_MAX_SINGLE_MESSAGE_CHARS=16000

# Assistant output cap (passed to provider as max_tokens)
# CHAT_MAX_TOKENS=2048
# Agent tool turns are shorter but may need longer wall time on local models.
# AGENT_TOOL_MAX_TOKENS=384
# When thinking is enabled, agent routing needs a much larger budget for reasoning.
# AGENT_TOOL_THINKING_MAX_TOKENS=2048
# AGENT_TOOL_TIMEOUT_SECONDS=90
# AGENT_EVIDENCE_BLOCK_MAX_CHARS=1200
# Local agent tool safety limits.
# AGENT_SHELL_TIMEOUT_SECONDS=30
# AGENT_TOOL_MAX_OUTPUT_BYTES=32768
# AGENT_FILE_MAX_BYTES=256000

# Fast parallel status text (Cursor-style live progress labels)
# Runs alongside search/agent work on a small model with thinking disabled.
# STATUS_TEXT_ENABLED=true
# STATUS_TEXT_PROVIDER=ollama
# STATUS_TEXT_MODEL=Qwen3-0.6B Q4_K_M
# STATUS_TEXT_BASE_URL=http://localhost:11434
# STATUS_TEXT_API_KEY=not-needed
# STATUS_TEXT_MAX_TOKENS=24
# STATUS_TEXT_TIMEOUT_SECONDS=1.5

# Attachment limits (mirrored in /api/health)
# ATTACHMENT_MAX_PER_MESSAGE=5
# ATTACHMENT_MAX_IMAGE_BYTES=5242880
# ATTACHMENT_MAX_TEXT_BYTES=524288

# Comma-separated model ids that accept image input (empty = keyword heuristic)
# MULTIMODAL_MODELS=qwen2-vl-7b

# Rapid-MLX serve performance (read by macos/serve.py when starting inference)
# RAPID_MLX_GPU_MEMORY_UTILIZATION=0.90
# RAPID_MLX_STREAM_INTERVAL=2
# RAPID_MLX_PREFILL_STEP_SIZE=8192
# RAPID_MLX_PREFILL_BATCH_SIZE=
# RAPID_MLX_COMPLETION_BATCH_SIZE=

# Exa web search (also loaded from ../Benchmarks/.env and ../ARS/.env if present)
# BENCHMARK_ENV_PATH=../Benchmarks/.env
# EXA_API_KEY=
# EXA_API_KEY_2=
# EXA_API_KEY_3=
# ARS_ENV_PATH=../ARS/.env
# EXA_SEARCH_TYPE=auto
# EXA_SEARCH_RESULTS=8
# EXA_CONTENT_MAX_AGE_HOURS=168
# EXA_BASE_URL=https://api.exa.ai
# EXA_KEY_STATE_PATH=~/.rs2/exa-key-state.json
# SEARCH_BUDGET_MS=30000
# SEARCH_EXA_TIMEOUT_MS=4000
# SEARCH_MEDIAWIKI_TIMEOUT_MS=4000
# MEDIAWIKI_USER_AGENT=RS2/1.0 (research)
# SEARCH_MAX_URLS=8
# SEARCH_FETCH_BUFFER=2
# SEARCH_FETCH_CONCURRENCY=4
# SEARCH_MAX_PASSAGES=12
# SEARCH_PASSAGE_CLIP_CHARS=600
# SEARCH_MAX_QUERY_FANOUT=6
# SEARCH_MAX_PAGE_CHARS=12000
# SEARCH_AGENT_MAX_ITERATIONS=10
# SEARCH_TOOL_BUDGET_MS=30000
# SEARCH_TOOL_MAX_URLS=4
# SEARCH_TOOL_MAX_PASSAGES=6
# SEARCH_TOOL_SOURCE_PASSAGE_CHARS=2400
# AGENT_TOOL_MAX_TOKENS=384
# AGENT_TOOL_TIMEOUT_SECONDS=90
# AGENT_EVIDENCE_BLOCK_MAX_CHARS=1200

# Firecrawl web search
# FIRECRAWL_API_KEY=
# FIRECRAWL_API_KEYS=
# FIRECRAWL_API_KEY_2=
# FIRECRAWL_KEY_STATE_PATH=~/.rs2/firecrawl-key-state.json
# FIRECRAWL_BASE_URL=https://api.firecrawl.dev
# FIRECRAWL_SEARCH_RESULTS=8
# SEARCH_FIRECRAWL_TIMEOUT_MS=15000

# Runtime pid/log directory (chat server + Rapid-MLX)
# RS2_RUNTIME_DIR=~/.cache/rs2

# Optional: load extra keys from sibling repos
# BENCHMARK_ENV_PATH=../Benchmarks/.env
# ARS_ENV_PATH=../ARS/.env
# HF_TOKEN_FILE=~/.thunder/hf-token
# EXA_API_KEYS=
# EXA_API_KEY=